What a token is
Models read and write in tokens, which are neither characters nor words. Frequent chunks collapse into one, so a common English word is a single token. Measured against o200k, English runs 0.20 tokens per character — about five characters each — and Korean 0.72. Billing is per token.
What counts as input is easy to miss
Input tokens are not just what the user typed this turn. The system prompt, any pasted documents, tool definitions, and in a chat the entire prior conversation all go back in on every turn. A ten-turn conversation sends far more input on the last turn than the first. This is the most common reason an estimate comes out too low.
Put the total of all of that into the input field above. For chat, multiplying by your average turn count lands closer to reality.
Korean runs about 1.3x
For the same content, Korean spends about 1.3x the tokens of English. Matching character counts makes it look like 3.65x, but the same content in Korean takes less than half the characters, so that comparison measures volume rather than language. The multiplier does not vary by model; what that 30% costs does, from ₩60 to ₩625,000 a month.
Read the breakdown across 225 models
It shifts again by tokenizer generation
The coefficients above follow o200k, the tokenizer used from the GPT-4o generation onward. With the previous generation's cl100k, Korean comes to 1.16 tokens per character, a factor of 1.6, while English does not move at all at 0.198. Older models cost more in Korean than the figures here suggest. Anthropic and Google publish no tokenizer, so there is no telling which side they sit on.
What this number leaves out
The rates here are the published list prices from the catalogue. Real bills are usually lower. Prompt caching discounts the repeated prefix, batch requests are close to half price at some vendors, and a few models have volume tiers. In the other direction, sending images or audio adds tokens on top.
Won figures use the exchange rate from the last collection. Contracts and invoices are in dollars, so leave room for the rate to move when you set a budget.
How far to trust these numbers
If you entered token counts yourself, the arithmetic is exact. If you pasted text, the estimate came within 7.6% by token weight across 926 paragraphs against o200k, though a single paragraph can be off by 48% and a word or two by more than double. Short strings are dominated by rounding, and it is long text that decides the bill, so real estimates are barely affected. Claude and Gemini publish no tokenizer, so expect more drift there. When you need precision, count once through the vendor's own token-counting API and put that number in the field above.